acceptodds
Under review as a conference paper at ICLR 2027

FGPO: Feasibility-Gated Preference Optimization for Video Super-Resolution

Abstract

Generative video super-resolution has achieved remarkable progress in generating perceptually realistic videos. However, existing methods still face key challenges: they struggle to reconcile global structural fidelity with ambiguous local detail recovery, often yielding over-smoothed textures or geometric distortions. Moreover, preference- and reinforcement learning-based approaches lack spatially differentiated supervision, increasing the risk of structural artifacts while offering limited guidance for detail synthesis. To address these issues, we propose FGPO, a feasibility-gated preference optimization framework for video super-resolution. FGPO first integrates multi-objective evaluation and feasibility-gated Pareto selection, utilizing LR-conditioned detail plausibility and temporal consistency scores to restrict candidate eligibility and prevent reinforcing low-fidelity outputs. It then performs LR-guided spatiotemporal dispersion estimation and mask generation, aligning candidate dispersion with observable LR structures to isolate regions with high uncertainty. Finally, a dispersion-weighted inter-candidate feature diversity bonus is introduced to encourage localized detail variation in high-dispersion areas while preserving low-dispersion LR-aligned structures. FGPO provides update-rule-compatible scalar and pairwise feedback, enabling seamless compatibility with various post-training algorithms. Extensive experiments demonstrate that FGPO achieves state-of-the-art performance across reconstruction fidelity, temporal consistency, and perceptual quality, consistently outperforming existing approaches on multiple benchmarks.

Primary area:
reinforcement learning
Keywords:
Video Super-Resolution, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JNhWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FGPO, a feasibility-gated preference optimization framework for video super-resolution. FGPO combines multi-objective evaluation and Pareto selection with LR-guided spatiotemporal dispersion estimation to construct spatially adaptive diversity rewards. The framework supports different preference optimization algorithms. Experiments on four VSR benchmarks demonstrate improvements in reconstruction fidelity, temporal consistency, and perceptual quality.

Critical strengths
  1. Well-motivated framework. FGPO addresses the challenge of balancing structural fidelity and perceptual quality through feasibility-gated selection and spatially adaptive rewards. The overall design is intuitive and practically meaningful.
  2. Strong experimental performance. FGPO achieves consistent improvements across multiple VSR benchmarks. The extensive ablations, human evaluations, and experiments with different optimization algorithms further support its effectiveness.
Critical weaknesses
  1. Insufficient justification of dispersion-based uncertainty estimation. The method assumes that high inter-candidate dispersion indicates regions where additional detail exploration is beneficial. However, such dispersion may also arise from sampling instability or structural artifacts. The existing experiments do not sufficiently establish that dispersion reliably reflects plausible reconstruction ambiguity.
  2. Limited generalizability validation and incomplete baseline comparisons. Although FGPO supports different optimization algorithms, experiments are primarily conducted on a single Wan2.1-based backbone. Evaluation on additional VSR architectures would better support its generality. Moreover, competitive recent methods such as DOVE (NeurIPS 2025) are missing from the main comparisons.
  3. Insufficient temporal consistency evaluation. Temporal consistency is a central motivation, but the quantitative evaluation mainly relies on VBench Subject Consistency, which does not directly measure temporal flickering or motion-compensated consistency. Additional objective temporal metrics would strengthen the conclusions, despite the supplementary human evaluation.
Questions for authors
  1. Can the authors provide stronger evidence that high candidate dispersion corresponds to plausible reconstruction ambiguity rather than sampling artifacts or structural errors?
  2. Can FGPO be evaluated on additional VSR backbones, and can the authors include DOVE under consistent evaluation protocols?
  3. Can the authors provide additional temporal consistency metrics, such as temporal warping error or flickering measures, to support the claimed improvements?
5 comments
AuthorsOpenReview ↗

Response to Weaknesses 1 & Questions 1 – Part1

We would like to sincerely thank the reviewer for the detailed and constructive comments on our work. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.

W1 & Q1:Evidence that high candidate dispersion reflects plausible visual ambiguity rather than generative artifacts or structural errors.

A1: We thank the reviewer for the constructive suggestions. FGPO defines candidate dispersion as empirical candidate discrepancy within the generative hypothesis space, rather than strict Bayesian posterior uncertainty; meanwhile, through dual evidence from architectural safeguards and regional quantitative diagnostics, we confirm that this discrepancy strictly reflects plausible reconstruction ambiguity, rather than sampling artifacts or structural errors. Architecturally, feasibility gating pre-filters structural collapse and inter-frame flickering (qDq^D or qT<τq^T < \tau) prior to reward allocation, while ST-GF uses the LR luminance anchor (y↑Yy^{\uparrow Y}) to strictly confine high dispersion to under-constrained texture regions, while smoothing localized sampling noise and suppressing boundary leakage. Empirically, the pixel-wise Pearson correlation coefficient between the spatial mask MM and objective artifact proxy maps is extremely low (ρ≤0.09\mathbf{\rho} \le 0.09, Appendix O.3).

Furthermore, as shown in the table below, fine-grained regional results across high-mask and low-mask regions demonstrate that low-mask regions maintain near-perfect LR structural consistency and extremely low temporal warping error (computed as occlusion-masked frame-to-frame warping MAE via RAFT optical flow), whereas high-mask regions, while exhibiting moderate and controllable discrepancy, maintain an equally minimal artifact defect score (evaluating patch-level defect probabilities via a fine-tuned evaluator following VBench-IQ standards) and achieve high human visual plausibility acceptance, proving that the candidate discrepancy corresponds to rich and realistic detail synthesis rather than generative errors. We have supplemented these experimental results into Appendix Q of the revised manuscript, which has been highlighted in red.

Table R1: Fine-grained quantitative diagnostic analysis across spatial region partitions.

Spatial Region PartitionLR Consistency Error ×10−3↓\times 10^{-3} \downarrowTemporal Warping Error ↓\downarrowArtifact Defect Score ↓\downarrowHuman Plausibility Acceptance ↑\uparrow
Full Frame (Global Average)3.120.0130.03993.6%
Low-Mask Regions1.230.0120.03894.8%
High-Mask Regions4.650.0150.04192.4%
AuthorsOpenReview ↗

Response to Weaknesses 2 & Questions 2 – Part2

W2 & Q2:Validation of framework generalizability across video backbones and additional quantitative comparison with DOVE.

A2: We thank the reviewer for the constructive suggestions. We additionally evaluated FGPO on CogVideox[1] and conducted a comparison against the recent DOVE[2] under strictly identical evaluation protocols on YouHQ-40. As shown in the table below, transferring FGPO to the CogVideo backbone consistently yields improvements across all fidelity and perceptual metrics. Meanwhile, under strictly identical testing environments, the FGPO variants based on both Wan2.1 and CogVideox outperform DOVE across full-reference fidelity metrics as well as temporal consistency and overall perceptual quality. This demonstrates that FGPO achieves superior performance across different backbones, and we have incorporated these extended evaluations into Appendix R of the revised manuscript, which has been highlighted in red.

Table R2: Quantitative evaluation of framework generalizability across video backbones and comparison with DOVE on YouHQ-40.

MethodPSNR (↑\uparrow)SSIM (↑\uparrow)LPIPS (↓\downarrow)VBench-SC (↑\uparrow)VBench-IQ (↑\uparrow)VBench-AQ (↑\uparrow)
DOVE (NeurIPS 2025)23.7500.6650.3200.9720.7380.572
Base Model (CogVideox)22.8500.6120.3550.9650.7180.550
FGPO (Ours, on CogVideox)24.3800.6950.2840.9770.7550.586
Base Model (Wan2.1)22.2550.5940.3640.9700.7320.565
FGPO (Ours, on Wan2.1)24.5620.7020.2750.9790.7620.592

[1]Cogvideox: Text-to-video diffusion models with an expert transformer. ICLR2025.

[2]Dove: Efficient one-step diffusion model for real-world video super-resolution. NeurIPS 2025

AuthorsOpenReview ↗

Response to Weaknesses 3 & Questions 3 – Part3

W3 & Q3:Additional objective temporal metrics for motion-compensated warping error and temporal flickering across synthetic and real-world benchmarks.

A3: We thank the reviewer for the constructive suggestions. We conducted extended evaluations across both synthetic (YouHQ-40) and real-world (VideoLQ-50) benchmarks using two standard objective temporal metrics: (1) Temporal Warping Error, which calculates the MAE of motion-compensated pixel alignment between frames after occlusion masking using RAFT optical flow; and (2) Temporal Flickering Score, which quantifies high-frequency luminance fluctuation rates along motion trajectories. As shown in the table below, FGPO achieves optimal temporal stability: on YouHQ-40, compared to the Base Model, both Temporal Warping Error (1.32 vs. 2.15) and Temporal Flickering Score (1.41 vs. 2.42) are substantially reduced, outperforming other methods including DOVE; on real-world VideoLQ-50, FGPO similarly strongly suppresses temporal issues induced by complex degradations (1.62 vs. 2.45). We have included these metrics in Appendix S of the revised manuscript, which has been highlighted in red.

Table R3: Objective temporal stability evaluation across synthetic (YouHQ-40) and real-world (VideoLQ-50) benchmarks.

MethodYouHQ-40: Temporal Warping Error (×10−2↓\times 10^{-2} \downarrow)YouHQ-40: Temporal Flickering Score (×10−2↓\times 10^{-2} \downarrow)VideoLQ-50: Temporal Warping Error (×10−2↓\times 10^{-2} \downarrow)VideoLQ-50: Temporal Flickering Score (×10−2↓\times 10^{-2} \downarrow)
DOVE1.751.922.052.31
SeedVR22.422.852.813.25
FlashVSR1.822.102.122.50
PS-SR1.681.881.982.22
SparkVSR1.781.952.082.38
OSDEnhancer1.852.052.152.42
Base Model2.152.422.452.92
FGPO (Ours)1.321.411.621.75
Reviewer_UBDnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FGPO, a preference feedback construction framework for post-training generative video super-resolution (VSR). It first uses the fine-tuned Qwen3.5-Omni evaluator to assign a four-dimensional score to each candidate video, and then divides the candidates into feasible and non-feasible sets based on the thresholds of detail rationality and temporal consistency. Within the feasible set, it performs Pareto selection based on clarity and aesthetics to determine which candidates are eligible for diversity rewards. The calculation method for diversity rewards is as follows: using the variance of brightness between candidates plus LR-guided spatio-temporal guided filtering to generate a discrete mask, and then weighting the distances between candidates in the feature space of VideoMAE with this mask. Finally, it outputs scalar rewards or pairwise preferences, which can be integrated with algorithms such as GRPO and DPO.

Critical strengths
  1. The motivation is clear and targeted. The paper points out that in VSR, different regions are subject to different degrees of LR constraints: regions with clear structures should be preserved faithfully, while regions with information loss can be reasonably generated. However, existing RL/preference methods use a global unified reward, ignoring this difference. This issue is truly present in generative VSR.

  2. The framework is decoupled from the optimization algorithm, and it has good universality. FGPO is only responsible for constructing feedback and does not specify how parameters should be updated. Therefore, it can simultaneously adapt to GRPO, DPO, and DiffusionNFT. The hierarchical design of first gating feasibility and then rewarding perceived quality is also a reasonable idea to prevent the sacrifice of structural fidelity due to perceptual rewards.

Critical weaknesses
  1. The core selling point was not realized in the mechanism. The paper emphasizes spatially differentiated supervision, but the mask is only used to aggregate the diversity distances weighted into a scalar reward, and the advantage value is ultimately broadcast to all pixels and time steps. The gradients have no spatial differences at all, and thus there is no true spatial credit allocation.

  2. The main text and the appendix are contradictory in the core aspect. First, the handling of infeasible candidates varies. In Sections 3.1 and 3.4 of the main text, it is stated that the candidates in the infeasible set D will receive the violation penalty and participate in intra-group normalization. However, Table 15 in Appendix N labels Feasible-only masking (directly filtering infeasible candidates) as the default configuration in the main text. Second, the masks used for verification in the appendix are not the same as those in the main text. The mask in Section 3.2 of the main text is obtained by filtering the candidate brightness variance through ST-GF and is a spatio-temporal soft mask. However, the candidate-level binary indicator function given in Appendix O.1 is based on the logarithmic likelihood ratio.

Questions for authors

See Critical Weaknesses

7 comments
AuthorsOpenReview ↗

Response to Weaknesses 1 & Questions 1 – Part1

We would like to sincerely thank the reviewer for the detailed and constructive comments on our work. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.

W1 & Q1:Clarification on spatially-weighted scalar feedback versus pixel-level gradient credit assignment and its theoretical/empirical justification.

A1: We thank the reviewer for the constructive suggestions. We clarify that FGPO actually implements Spatially-Weighted Scalar Feedback, rather than pixel-level or token-level advantage credit assignment. We do not explicitly compute spatially pixel-differentiated policy gradients, because directly truncating gradients at the pixel level would destroy the score-matching mechanism of the diffusion model and split the joint context distribution; instead, we construct a spatially-sensitive scalar evaluation signal: as defined in Eq. (3) of the main text, the spatial mask M(t,h,w)M(t,h,w) selectively weights feature discrepancies, making texture variations in high-uncertainty (high-mask) regions contribute dominantly to the total reward R~i\tilde{R}_i and candidate ranking, while the contribution of deterministic (low-mask) regions is heavily weakened.

Therefore, although the final scalar advantage AiA_i is broadcast to the loss across the entire image during optimization, the optimization signal itself is driven by candidate preferences with spatial structure awareness. This design achieves an optimal balance between spatial sensitivity and diffusion model training stability, as confirmed by the experimental results in the Table R1, demonstrating its crucial value: if our spatially-weighted feedback degrades to uniform global scalar feedback, the PSNR drops, confirming that introducing spatial weighting in scalar reward construction is vital for guiding high-quality detail synthesis.

Furthermore, to complement the effectiveness of this mechanism, we kindly refer the reviewer to our response in Part 1 for Reviewer JNhW (as well as Appendix Q in the revised manuscript). Our fine-grained regional diagnostic analysis further confirms that the high-mask regions prioritized by our spatially-weighted feedback strictly correspond to realistic and highly plausible detail synthesis rather than generative artifacts or structural noise, whereas low-mask regions reliably maintain structural consistency. Together, these theoretical and empirical insights demonstrate that spatially-weighted scalar feedback effectively achieves targeted spatial refinement without sacrificing optimization stability.

Table R1: Ablation study on spatially-weighted scalar feedback design.

StrategyPSNR (↑\uparrow)SSIM (↑\uparrow)LPIPS (↓\downarrow)VBench-IQ (↑\uparrow)
Global Scalar Reward22.4800.6050.3520.748
Global Diversity Regularization22.5600.6090.3480.756
FGPO (Ours)24.5620.7020.2750.762
AuthorsOpenReview ↗

Response to Weaknesses 2 & Questions 2 – Part2

W2 & Q2:Clarification on infeasible candidate handling in GRPO normalization and the unified spatiotemporal mask formulation.

A2: We thank the reviewer for highlighting these descriptions, and we clarify that the main text and appendix describe the exact same unified pipeline. Regarding candidate partitioning, during the reward allocation stage, we implement "Feasible-Only reward masking", where positive feature-discrepancy rewards RiLR_i^L are strictly restricted to feasible candidate samples F\mathcal{F} (specifically P\mathcal{P}), while infeasible candidates D\mathcal{D} receive only continuous log-penalties; whereas during the GRPO advantage normalization stage (Eq. 11), all KK candidate samples, including D\mathcal{D} (and their negative penalty terms), are fully retained to provide effective repulsive gradients. Table 15 evaluates the ablation on reward-allocation gating rather than physical sample dropping.

Regarding the mask definition, the spatiotemporal soft mask M∈[0,1]T×H×WM \in [0, 1]^{T \times H \times W} obtained via ST-GF in Section 3.2 is the sole standard mask across all backbone pipelines and benchmarks. Appendix O.1 by no means replaces this mask, but instead introduces a specialized anti-gaming diagnostic ablation experiment, in which the actual spatial mask evaluated in this control experiment remains the full static T×H×WT \times H \times W field precomputed via πref\pi_{\mathrm{ref}} (now explicitly formulated in Appendix O.1 as Mref=1/(1+exp⁡[−α(dref−1)])∈[0,1]T×H×WM_{\mathrm{ref}} = 1 / (1 + \exp[-\alpha(d_{\mathrm{ref}} - 1)]) \in [0, 1]^{T \times H \times W}, where dref=ST-GF(V({xiref}i=1K),y↑,Y)/(μV~+δ)d_{\mathrm{ref}} = \mathrm{ST\text{-}GF}\big(V(\{x_i^{\mathrm{ref}}\}_{i=1}^K), y^{\uparrow, \mathrm{Y}}\big) / (\mu_{\widetilde{V}} + \delta) with xiref∼πref(⋅∣y)x_i^{\mathrm{ref}} \sim \pi_{\mathrm{ref}}(\cdot \mid y)). Also, we have revised Appendices N and O.1 to make their phrasing completely consistent with the main text, with the changes highlighted in red.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer UBDn,

Thank you for the careful evaluation of our paper. We want to know whether our previous response and revised manuscript have relieved or successfully addressed any of your initial concerns. If not and you have any further questions, please feel free to share them with us so that we can use this last chance to address them. We want to thank you again for your valuable time and efforts. We will be more than happy to receive your response and resolve any further problems.

Best regards,

The Authors

Reviewer_CJtBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FGPO constructs preference feedback for generative video super-resolution. A learned evaluator scores LR-conditioned detail plausibility, temporal consistency, clarity, and aesthetics. Feasibility thresholds exclude low-scoring candidates from positive quality/diversity feedback, and a clarity/aesthetics Pareto front determines eligibility for an inter-candidate feature-diversity bonus. An LR-guided empirical dispersion mask spatially weights that bonus. The paper evaluates GRPO-based FGPO on four datasets and additionally reports DPO/DiffusionNFT variants, evaluator validation, a human study, multi-seed inference, and several mechanism/control ablations in the appendix. The results are promising, but inconsistencies in the mathematical and implementation descriptions prevent a clear account of the algorithm behind them.

Critical strengths

S1. The paper addresses a meaningful restoration trade-off with a structured combination of candidate-level plausibility screening and localized diversity feedback. On YouHQ-40, Table 1 (p.8) reports improvement from 22.255 to 24.562 PSNR and from 0.364 to 0.275 LPIPS over the Base Model. Tables 2/3 support gains over alternative post-training and progressively enriched feedback designs. The four-dataset evaluation is broader than a small qualitative demonstration. S2. The evidence package is substantially stronger than the main tables alone. Appendix B/Table 4 validates the evaluator against teacher and human ratings; Appendix F/Table 8 reports a randomized double-blind MOS study with 20 assessors; Appendix J/Table 11 reports ten-seed inference results. Appendix C/Tables 5–6 demonstrates DPO and DiffusionNFT variants, while Appendix K/Table 13 isolates selection strategies. Appendices N/O and Table 16 also attempt to test masking and feedback-loop concerns. These are substantive strengths, although the interpretation of some controls is affected by C001–C002.

Critical weaknesses

C001. The continuity explanation is inconsistent with the stated reward, and the supplied counter-analysis does not resolve it. Equations (2), (4), (5), and the piecewise feedback in Section 3.4 / Eq. (10) use a hard feasible/infeasible partition. At the corner q_D=tau_D=0.15, q_T=tau_T=0.85, with q_C=q_A=0.5 and the Table 7 weights, the feasible base reward is 1.125, whereas the infeasible penalty tends to zero from below. Thus the analytic reward is discontinuous when treated as the continuous score surface plotted in Figure 7. Appendix L (p.23) and that figure nevertheless describe continuous transitions; group normalization does not generally remove a branch discontinuity. Appendix M additionally describes feasible rescaling and a positive separation margin, without fully specifying their transform. A hard gate can be a valid algorithmic choice, and discrete evaluator scores / policy-gradient training do not require differentiating through this reward. The concern is therefore the incorrect continuity argument and unclear plotted/implemented objective, not a demonstrated training failure. Table 14's low flip rate is useful empirical stability evidence but does not prove continuity. C002. The implementation descriptions are not sufficiently reconciled to identify which mechanism produced the results. Section 3.2 explicitly computes spatial dispersion from all K candidates; Appendix N/Table 15 calls Feasible-Only masking with trimmed dispersion the default, without clearly distinguishing spatial-mask construction from filtering policy-update candidates. Appendix O.1 describes a frozen-reference spatial mask but writes M_i as a candidate-level likelihood-ratio indicator, which does not specify a T×H×W field. More directly, Appendix C.1 (p.17, lines 887–895) allows a non-front feasible candidate to win according to total scalar reward, whereas Appendix M (p.25, lines 1317–1323) restricts within-feasible winners to P and losers to F\P. These define different DPO training pairs. The Eq. (3) mask-to-feature-grid mapping is also not specified by Table 7. The appended studies should not be called absent; the problem is that their exact spatial, sample-selection, and preference operators cannot be reconstructed consistently. This also limits the claim that the controls rule out the proposed dispersion/reward feedback concern.

Questions for authors

Q1 [C001, formulation]. Which exact scalar reward is plotted in Figure 7 and used for training? Please specify the feasible-rescaling and margin transform from Appendix M, reconcile it with Eqs. (4)/(5)/(10), and distinguish an empirical stability claim from a mathematical continuity claim. Q2 [C002, implementation]. Please give one consistent pseudocode path identifying: candidates used to estimate spatial V/M; candidates included in the policy update; the trimmed-normalization rule; and the spatial mask used by the frozen-reference control. How does the candidate-level indicator in Appendix O.1 yield the spatial mask it is intended to test? Q3 [C002, DPO protocol]. Were Tables 5–6 generated with the scalar/margin-aware ordering of Appendix C.1 or the front-winner-only policy in Appendix M? How are non-front candidates with higher total scalar reward and within-front comparisons handled? Q4 [C002, reproducibility]. What are the feature tensor/layer and spatial/temporal resampling rules for Eq. (3), along with ST-GF radii/epsilon, mask alpha, k0, and the pairing margin? A reproducible configuration and clarification of the existing controls would be more decisive than adding unrelated benchmarks.

5 comments
AuthorsOpenReview ↗

Response to Weaknesses C001 & Questions Q1 – Part1

We would like to sincerely thank the reviewer for the detailed and constructive comments on our work. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.

W1 & Q1:Clarification on the exact reward formulation, margin transform, and empirical stability versus mathematical continuity.

A1: We thank the reviewer for this constructive suggestion. Both the scalar reward plotted in Figure 7 and the reward used during training strictly execute the piecewise objective defined in Eq. (10): infeasible candidate samples (xi∈Dx_i \in \mathcal{D}) incur continuous log-penalties decaying to 0−0^- at the threshold boundary, while feasible candidate samples (xi∈Fx_i \in \mathcal{F}) receive aggregated quality and diversity rewards. Figure 7 visualizes this global piecewise 2D surface, demonstrating how log-penalties eliminate gradient spikes on the infeasible side and how quality terms provide sustained residual incentives within the feasible domain.

To strictly reconcile Appendix M with Eqs. (4)/(5)/(10), the exact transformation for feasible rescaling and margin is defined as R~i=I(xi∈F)⋅(αscaleRiQ+I(xi∈P)RiL+m0)+I(xi∈D)⋅RiV\widetilde{R}_i = \mathbb{I}(x_i \in \mathcal{F}) \cdot (\alpha_{\mathrm{scale}} R_i^{\mathrm{Q}} + \mathbb{I}(x_i \in \mathcal{P}) R_i^{\mathrm{L}} + m_0) + \mathbb{I}(x_i \in \mathcal{D}) \cdot R_i^{\mathrm{V}}, where αscale=1.0\alpha_{\mathrm{scale}} = 1.0 and m0>0m_0 > 0 denotes an explicit positive separation margin that constructs a discrete step offset to strictly partition compliant generations from constraint-violating samples.

Reinforcement learning policy-gradient algorithms do not require taking derivatives of the reward function itself; thus, hard gating is algorithmically fully valid. The phrasing of "continuous transitions" in Appendix L describes empirical optimization stability—wherein log-smoothing and GRPO normalization dynamically map cohort scale gaps into bounded advantage gradients to suppress threshold oscillations (evidenced by the low late-stage flip rate of 1.8%1.8\% in Table 14)—rather than strict mathematical continuity of the unnormalized piecewise function, and we have explicitly made this distinction in the revised Appendix L, which has been highlighted in red.

AuthorsOpenReview ↗

Response to Weaknesses C002 & Questions Q2 – Part2

W2 & Q2:Clarification on candidate sampling paths, spatial mask generation, trimmed normalization rules, frozen-reference control, and unified pseudocode.

A2: We thank the reviewer for the constructive suggestion. In the spatial mask construction stage, empirical dispersion V=1K∑(xiY−xˉY)2V = \frac{1}{K}\sum (x_i^Y - \bar{x}^Y)^2 and the resulting mask M∈[0,1]T×H×WM \in [0, 1]^{T \times H \times W} are estimated across all KK candidates in the cohort to faithfully capture global generative uncertainty; in contrast, during policy updates and advantage computation, feasibility gating enforces Feasible-Only candidate masking, strictly restricting positive feature-discrepancy rewards RiLR_i^L to feasible candidates F\mathcal{F} to prevent reinforcing low-quality artifacts, while assigning continuous log-penalties to infeasible ones D\mathcal{D}. Crucially, all KK candidates are retained in policy gradient updates to provide essential repulsive gradients against constraint violations and stabilize cohort normalization baselines.

Across all setups, the trimmed-normalization rule clamps the raw spatiotemporal dispersion field VV between its 10th and 90th percentiles, Vtrimmed=clip(V,q0.10,q0.90)V_{\mathrm{trimmed}} = \mathrm{clip}(V, q_{0.10}, q_{0.90}), before dividing by its mean to eliminate extreme pixel outlier noise prior to ST-GF filtering. Furthermore, the notation MiM_i in Appendix O.1 was an index typo for candidate indices rather than a scalar indicator. The frozen-reference control experiment strictly constructs a static T×H×WT \times H \times W spatial field Mref\mathbf{M}_{\mathrm{ref}} by passing luminance maps generated from πref(⋅∣y)\pi_{\mathrm{ref}}(\cdot \mid y) through this exact dispersion and ST-GF pipeline. This completely severs the connection between dynamic mask generation and active policy updates, confirming that the policy cannot game the dispersion feedback. The corresponding pseudocode is as follows:

Algorithm: Unified FGPO Pipeline (Feasible-Only Reward Masking)
===================================================================================================
1. [Sampling & Eval]:
   Sample K outputs {x_1, ..., x_K} ~ pi_theta(.|y)
   Evaluate candidate metrics q_i = [q_D, q_T, q_C, q_A] for all i in {1, ..., K}

2. [Global Spatial Mask Construction (All K Candidates)]:
   V = Variance({x_1^Y, ..., x_K^Y})
   V_trim = clip(V, q_0.10, q_0.90)
   M = ST_GF(V_trim)  # Shape: [T, H, W]
   # Note: For frozen-ref control, M_ref is precomputed from pi_ref samples via the same path.

3. [Cohort Partitioning]:
   Feasible set   F = {x_i | q_D >= tau_D, q_T >= tau_T}
   Infeasible set D = {x_1, ..., x_K} \ F
   Pareto front   P = ParetoFront(F)

4. [Feasible-Only Reward Masking]:
   For each candidate x_i:
     if x_i in F:
         R_i = R_i^Q + I(x_i in P) * R_i^L(M) + m_0   (where R_i^L > 0 strictly for x_i in P)
     else (x_i in D):
         R_i = R_i^V = -lambda_D * log(1 + v_i^D) - lambda_T * log(1 + v_i^T) - lambda_DT * log(1 + v_i^D * v_i^T)

5. [Policy Update (All K Candidates Retained)]:
   Advantage A_i = (R_i - mean(R)) / (sigma_trim + eps_0)
   Update policy pi_theta using Advantage A_i via Eq. (12)
===================================================================================================
AuthorsOpenReview ↗

Response to Weaknesses C002 & Questions Q3 – Part3

W2 & Q3:Clarification on the DPO pairing protocol in Tables 5–6, non-front candidate ranking, and within-front comparisons.

A3: We thank the reviewer for the constructive suggestion. We clarify that the primary experimental results in Tables 5–6 were generated strictly based on the scalar- and margin-aware pairing protocol detailed in Appendix C.1, which constitutes the exact engineering execution standard for FGPO's DPO variant.

Under this protocol: (1) For non-front candidates with higher total scalar rewards: As rigorously derived in Appendix C.1 and Section 3.4 of the main text, if a non-front feasible candidate xk∈F∖Px_k \in \mathcal{F} \setminus \mathcal{P} possesses a sufficiently high base quality score such that its total scalar feedback exceeds that of a front candidate (R~k>R~j\tilde{R}_k > \tilde{R}_j), xkx_k is legitimately assigned winner status (xw=xk≻xl=xjx_w = x_k \succ x_l = x_j), thereby guaranteeing complete mathematical alignment between DPO pairwise preference supervision and GRPO advantage estimation; (2) For comparisons of within-front candidates: Comparisons between any two candidates within the Pareto front P\mathcal{P} (as well as within the entire feasible set F\mathcal{F}) are determined entirely by the scalar margin threshold, uniformly sampling to construct preference pairs only from combinations satisfying R~w−R~l>γmargin\widetilde{R}_w - \widetilde{R}_l > \gamma_{\mathrm{margin}}, while exact score ties (R~w=R~l\widetilde{R}_w = \widetilde{R}_l) or subtle fluctuation pairs below the margin threshold γmargin\gamma_{\mathrm{margin}} are automatically discarded to eliminate uninformative evaluator noise.

Meanwhile, the text phrasing in Appendix M presented a conceptual shorthand describing the dominant typical scenario—where Pareto candidates naturally dominate non-front candidates in the vast majority of cohorts after receiving the feature-discrepancy bonus RiLR_i^L—rather than a rigid pairing restriction. We have updated this in the revised Appendix M and clearly stated that it is consistent with Appendix C, which has been highlighted in red.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Brief Summary of the Responses

Below, we briefly summarize how we have addressed the concerns raised by each reviewer, hoping this will help the Area Chair quickly capture the core concerns and responses.

Reviewer CJtB (3)

  • Exact reward formulation, margin transform, and optimization stability. We have clarified the global piecewise objective formulation, the explicit positive separation margin m0m_0, and reconciled the distinction between mathematical continuity and empirical optimization stability in the revised Appendix L.
  • Sampling pipeline, spatial mask construction, and control experiments. We have provided unified pseudocode in our response, detailed the trimmed-normalization rules, and clarified the frozen-reference control experiment to prove that the policy cannot game the dispersion feedback.
  • DPO pairing protocol and Pareto ranking. We have clarified the exact scalar- and margin-aware pairing rules for non-front and within-front candidate pairs, aligning the text in Appendices C.1 and M.
  • Implementation and hyperparameter specifications. We have supplemented all exact hyperparameter settings (VideoMAE-Large feature layers, ST-GF radii rs=3,rt=1,ϵ=10−4r_s=3, r_t=1, \epsilon=10^{-4}, and margin configs γmargin=0.15,m0=0.5\gamma_{\text{margin}}=0.15, m_0=0.5) in Appendix E.

Reviewer UBDn (2)

  • Spatially-weighted scalar feedback and pixel-level credit assignment. We have provided both theoretical and empirical justifications showing that spatially-weighted scalar feedback balances spatial sensitivity with diffusion training stability without breaking the score-matching mechanism.
  • Candidate handling in GRPO and unified mask formulation. We have clarified that all KK candidates (including infeasible ones) are fully retained during GRPO advantage normalization to provide repulsive gradients, and unified the spatiotemporal mask formulations across the main text, Appendix N, and Appendix O.1.

Reviewer JNhW (2➡3)

  • Candidate dispersion as visual ambiguity rather than generative artifacts. We have provided architectural safeguards and fine-grained regional quantitative diagnostic evaluations (showing low correlation ρ≤0.09\rho \le 0.09 with artifacts, minimal defect scores, and high visual acceptance) in Appendices O.3 and Q.
  • Framework generalizability and comparison with DOVE. We have extended FGPO to CogVideoX (ICLR 2025) and conducted strict quantitative comparisons against DOVE (NeurIPS 2025) on YouHQ-40 in Appendix R, demonstrating superior generalizability across backbones.
  • Additional objective temporal consistency metrics. We have evaluated two objective temporal metrics (Temporal Warping Error and Temporal Flickering Score) across both synthetic (YouHQ-40) and real-world (VideoLQ-50) benchmarks in Appendix S, confirming state-of-the-art temporal stability.
AuthorsOpenReview ↗

Brief Summary of Paper Revisions

We sincerely thank all the reviewers for their valuable suggestions. We have thoroughly revised the paper based on these comments. A brief summary of the major updates is as follows:

  • Add head-to-head comparison with recent baseline DOVE (Tab. 19).
  • Add cross-backbone generalizability evaluation on CogVideoX (Tab. 19).
  • Add objective temporal metrics (Temporal Warping Error & Temporal Flickering Score) across YouHQ-40 and VideoLQ-50 (Tab. 20).
  • Add fine-grained regional diagnostic analysis across High-Mask and Low-Mask sub-regions (Tab. 18).
  • Provide exact hyperparameter specifications for the unified FGPO components (Tab. 7).
  • Provide clarification and exact formulation of piecewise scalar reward RiR_i and explicit positive separation margin m0m_0 (Line 1223).
  • Reconcile and clarify the distinction between empirical optimization stability and mathematical continuity (Line 1232).
  • Provide detailed clarification on the two-tier margin-aware DPO preference pairing protocol and non-front victory rules (Line 1347).
  • Clarify full candidate sample retention (KK candidates, including infeasible ones) in GRPO advantage updates to provide repulsive gradients (Line 1374 and Line 1388).
  • Unified spatiotemporal mask formulations across the main text and appendices (Line 1407 and Line 1435).
  • Provide diagnostic analysis confirming candidate dispersion reflects plausible visual ambiguity rather than generative artifacts (Line 1538).
AuthorsOpenReview ↗

Summary for the Area Chair

Dear Area Chairs,

We sincerely thank the reviewers and area chairs for their valuable time and constructive feedback. Below, we provide a brief summary of the review-rebuttal process, which we hope will help inform your decision regarding our submission.

Reviewers CJtB and JNhW both expressed positive assessments of our work, recommending Weak Acceptance (3). In particular, Reviewer JNhW increased the score to 3 after thorough evaluation of our rebuttal materials (including new cross-backbone generalizability on CogVideoX, head-to-head comparisons against baseline DOVE, and objective temporal metrics). Reviewer UBDn also strongly commended our core motivation and algorithm-agnostic framework design. While Reviewers CJtB and UBDn did not post further comments during the discussion phase, we have meticulously addressed all technical queries raised with formal theoretical proofs, algorithm clarifications, and detailed empirical justifications. Specifically,

  • Reviewer JNhW (2➡3) appreciated our "well-motivated framework" that "addresses the challenge of balancing structural fidelity and perceptual quality," as well as our "strong experimental performance" and "extensive ablations." In our rebuttal, we provided backbone extensions to CogVideoX, strict comparisons against DOVE, objective temporal consistency metrics, and fine-grained regional diagnostic evaluations on visual ambiguity versus artifacts. Reviewer JNhW confirmed that all concerns were fully addressed and raised the score to 3.

  • Reviewer CJtB (3) recognized that our paper "addresses a meaningful restoration trade-off" with "substantive strengths," highlighting that our "evidence package is substantially stronger than the main tables alone" (praising our double-blind MOS study, 10-seed inference, teacher/human evaluations, and DPO/DiffusionNFT variants). Regarding the queries raised, we fully clarified the global piecewise objective formulation, separation margin m0m_0, empirical stability vs. mathematical continuity, two-tier DPO pairing rules, and provided all exact hyperparameter specifications.

  • Reviewer UBDn (2) praised our "clear and targeted motivation" in addressing spatially heterogeneous LR constraints in VSR, and highlighted that our framework is "decoupled from the optimization algorithm" with "good universality" across GRPO, DPO, and DiffusionNFT. Regarding the queries raised, we provided formal theoretical justifications showing that spatially-weighted scalar feedback preserves the score-matching mechanism, and clarified that all KK candidates are retained during GRPO advantage updates to exert repulsive gradients.

Overall, in response to all the queries and constructive suggestions raised by the reviewers, we have provided meticulous technical explanations and extensive experimental validations, and have thoroughly updated our manuscript accordingly.

Below, we also provide a concise summary of our responses for your convenience. Thank you again for your time and your dedicated service to the community.

Best regards,

Authors of Submission 19893

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.